Disks and files
Disk and branch names can use letters, digits,
-, and _.
Disks per plan
Disks synchronized with a data source
A disk synchronized with a data source keeps each file as a normal object in your bucket, so these limits apply instead.
The path length includes the bucket prefix, if the disk uses one.
Throughput and requests
Latency
Checkpoint and branch times don’t depend on disk size. A client cache miss is served from the file server’s
cache. Data that has to come from the data source takes longer, typically 10 to 30 ms for Amazon S3.
An
O_SYNC write returns once the file server has made it durable. We measured these on October 27, 2026 from a c7gn.2xlarge in AWS us-east-1, using the script in
Reproduce these numbers.
Sandboxes
A sandbox pauses when its session reaches 24 hours or its idle timeout.
Serverless execution
S3 API
The last part of a multipart upload can be smaller than 5 MiB.
Reproduce these numbers
You can measure every latency row yourself with one script that uses the public API, thearchil CLI, and fio.
Run it in the same cloud region as the disk, since a cross-region round trip is larger than most of these figures.
Script and our results
Script and our results
It needs Our run, from a c7gn.2xlarge in AWS
curl, jq, fio, journalctl, the archil CLI, and an API key
in ARCHIL_API_KEY. Set API to your region’s control plane endpoint and
REGION to match, then run it as root with sudo -E bash latency.sh. It takes about two minutes, creates disks named
latency-test-*, and deletes them when it finishes.latency.sh
us-east-1 on October 27, 2026:- Create and delete leave out the TLS handshake, about 10 ms on a new connection.
mount ms, wallincludes starting the CLI. The client log line doesn’t.- Checkpoint and branch are CLI wall times, so they include starting the CLI, and branch also opens a new TLS connection.
- The read test mounts with
--read-ahead 0. With the default, random reads also pull in neighboring blocks, so about a third of them hit the client cache.